acceptodds
Withdrawn from ICLR 2027

DTNO: Direct-Time Neural Operators for One-Pass Spatiotemporal Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Autoregressive neural PDE surrogates advance a system one or a few steps at a time, causing inference cost to grow with forecast horizon and allowing prediction errors to accumulate through repeated feedback. We study the Direct-Time Neural Operator (DTNO), a finite-time neural operator that predicts a requested future state in a single network evaluation. DTNO decomposes forecasting into a horizon-independent encoder, a lead-time-conditioned latent evolution operator, and a residual decoder. Lead time modulates the Fourier operator blocks through feature-wise linear modulation, enabling a single model to predict across multiple forecast horizons without autoregressive rollout. We evaluate DTNO on five spatiotemporal systems ranging from reaction-diffusion dynamics to reacting flows, using a common FNO backbone, matched training splits, and the same evaluation protocol as an autoregressive counterpart. For a requested future state, DTNO uses a single forward pass rather than a number of rollout steps that grows with the forecast horizon. Across these systems, autoregressive forecasting is more accurate at short lead times, whereas DTNO consistently achieves superior accuracy at medium horizons, with accuracy transition points emerging between \(h=8.1\) and \(h=19.1\). At \(h=32\), DTNO achieves lower relative \(L_2\) field error than the matched autoregressive baseline on all five datasets, while eliminating recursive error feedback and requiring no sequential rollout between the observed state and the queried horizon. DTNO also exhibits substantially lower variation across repeated training runs. These results show that direct finite-time prediction provides a practical alternative to autoregressive rollout for medium- and long-horizon spatiotemporal forecasting. By conditioning a shared neural operator directly on forecast time, DTNO combines horizon-flexible prediction, constant-depth inference, and improved long-horizon stability within a single model.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Neural Operators, Operator Learning, Spatiotemporal Forecasting, Dynamical Systems, Partial Differential Equations

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_urbMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies Direct-Time Neural Operators (DTNO) for predicting a future PDE state from an observed history and a requested lead time in a single forward pass. The implementation combines an FNO backbone, Fourier time embeddings, FiLM conditioning, and residual decoding, with an optional latent semigroup regularizer. Across five datasets, the authors compare direct prediction with a separately trained autoregressive counterpart and examine accuracy–horizon trade-offs, inference cost, seed variability, and area-relative performance. DTNO outperforms the autoregressive FNO baseline at intermediate horizons, although its long-horizon predictions often approach long-term average and its advantage depends substantially on the backbone.

Critical strengths
  1. The paper addresses the accuracy and computational cost of long-horizon forecasting for spatiotemporal systems, an important problem in scientific machine learning.
  2. The proposed model uses a straightforward architecture to support predictions at multiple lead times within a single model, with one forward pass per requested future state.
  3. The experiments cover five datasets including reaction–diffusion, fluid dynamics, and reacting flows, providing a reasonably comprehensive evaluation of the approach
Critical weaknesses
  1. The methodological contribution is limited, and the value of the particular combination is not established. The method combines several existing components: an FNO (Li et al., 2021), Fourier time embeddings (Xu et al., 2019, Tancik et al., 2020), FiLM (Perez et al., 2018), residual connections, and an optional consistency loss. Time-conditioned direct prediction is already central to Poseidon (Herde et al., 2024) and TC-UNet (Ovadia et al., 2025), as the manuscript itself acknowledges in Section 2. Furthermore, Section 3.2 states that the pointwise encoder and decoder contain little capacity, so conditioning the propagator effectively conditions the entire network. It is therefore unclear what substantive benefit follows from the proposed factorization. The reported ablations do not isolate its advantage over simpler time conditioning, or establish the contributions of the history window and residual decoder. The main comparison establishes a difference between temporal prediction strategies, but offers limited evidence for DTNO-specific design choices.

  2. The long-horizon results provide limited evidence of useful dynamical prediction. On several datasets, DT-FNO performs similarly to or worse than time-averaged reference predictions (Tables 2 and 8). Although these references use test-trajectory information, the results raise questions about whether DT-FNO preserves meaningful dynamics beyond avoiding rollout instability. Stronger validation on challenging, non-periodic systems would help establish its practical forecasting value.

  3. The connection between the latent semigroup constraint and physical evolution is insufficiently justified. Equation (3) encourages Φ(z,0)=z\Phi(z,0)=z, but substituting this into Equation (2) gives Gθ(Xt,0)=U(t)+Dθ(Eθ(Xt)),G_\theta(X_t,0)=U(t)+D_\theta(E_\theta(X_t)), which does not ensure the required zero-time identity unless the decoded increment vanishes. Likewise, latent composition consistency does not by itself establish consistency of the decoded physical predictions. The manuscript does not establish that Φ(E(Xt),τa)\Phi(E(X_t),\tau_a) represents the encoding of an appropriately advanced history, and the residual decoder remains referenced to the original observed state. These issues weaken the interpretation of the regularizer as enforcing physical temporal structure. Moreover, increasing latent RMS rules out collapse to zero, but does not rule out loss of sample-dependent information or other degenerate representations as shown in experimental section. Consequently, the null accuracy result is difficult to interpret as evidence about the usefulness of physically meaningful semigroup consistency.

  4. The reported results contain inconsistencies. Tables 2 and 9 report different RealPDEBench results for DT-FNO and SG-DT-FNO without explaining the discrepancy. Section 5.3 also claims that DT-FNO outperforms the (R=16) autoregressive model at every tabulated horizon, yet Table 3 includes a Gray–Scott autoregressive seed with lower one-step error. These inconsistencies should be clarified to support the experimental conclusions.

  5. Table 4 contains the placeholder “[place tab_physics.tex next to main.tex]” or improper importing tex file.

References

  • Li, Z., et al. “Fourier Neural Operator for Parametric Partial Differential Equations.” ICLR, 2021.
  • Perez, E., et al. “FiLM: Visual Reasoning with a General Conditioning Layer.” AAAI, 2018.
  • Herde, M., et al. “Poseidon: Efficient Foundation Models for PDEs.” NeurIPS, 2024.
  • Ovadia, O., et al. “Real-time inference and extrapolation with Time-Conditioned UNet: Applications in hypersonic flows, incompressible flows, and global temperature forecasting.” Computer Methods in Applied Mechanics and Engineering.
  • Xu, D., et al. “Self-attention with Functional Time Representation Learning.” NeurIPS, 2019.
  • Tancik, M., et al. “Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains.” NeurIPS, 2020.
Questions for authors
  1. What measurable benefit does the proposed encoder–propagator–decoder factorization provide over a conventional time-conditioned FNO? Can the authors isolate the effects of FiLM versus simpler conditioning, the history window, and residual decoding?
  2. What hypothesis explains the crossover and its disappearance with a U-Net? Can the authors test it through a controlled within-dataset change in sampling cadence or dynamical difficulty?
  3. How does Equation (3) imply zero-time identity or composition consistency for the physical prediction in Equation (2)? Can the authors evaluate decoded consistency and alignment with the encoding of an advanced history, beyond latent RMS?
  4. Can the authors provide the missing Table 4, explain the different values in Tables 2 and 9, and correct the statement that the (R=16) direct model achieves better performance at every tabulated horizon?
3 comments

Thank you for the detailed review. We clarify the contribution and acknowledge the mathematical and reporting limitations below.

1. Contribution and factorization.

Our contribution is an empirical comparison of separately trained direct and autoregressive models with matched spatial backbones, trajectory-held-out splits, and horizon-wise evaluation. We acknowledge prior time-conditioned methods, including Poseidon's comparison of inference modes.

The encoder–propagator–decoder factorization is an implementation choice, not a new operator class. Our ablations do not establish separate gains from FiLM, history length, or residual decoding. The contribution we defend is the controlled comparison and its boundaries, including backbone dependence and sparse-horizon interpolation failures. Missing external direct-time baselines limit broader comparative claims.

2. Useful prediction versus long-horizon averaging.

We agree that bounded error does not establish meaningful dynamics; Section 5.5 and Table 8 explicitly document this limitation. Nevertheless, Table 2 reports intermediate-horizon benefits: on Gray–Scott at h=32, DT-FNO has field error 0.3683, versus 0.5882 for AR-FNO-R and 0.5639 for trajectory climatology.

This supports field prediction at that horizon, not sustained chaotic trajectory tracking. Cylinder also has a favorable interval, but its quasi-periodicity limits that interpretation. The test-trajectory mean is an oracle diagnostic, not a deployable predictor or a rigorous information limit. Similarity to it does not prove that the observations contain no remaining predictive information.

3. Crossover and the U-Net result.

A plausible explanation is that multihorizon direct training trades some short-horizon accuracy for later-horizon performance, while autoregressive prediction accumulates approximation error through model-generated inputs. The disappearance of the advantage with an accurate, stable U-Net rollout is consistent with this hypothesis.

However, the width, mode-count, and precision controls do not establish a unique causal mechanism. Appendix D.5 provides no completed controlled cadence test. The 8–20-step crossover is therefore an observation for the tested FNO settings, not a universal physical timescale.

4. Latent versus physical consistency.

Your mathematical objection is correct. Even if Φθ(Eθ(Xt),0)=Eθ(Xt)\Phi_\theta(E_\theta(X_t),0)=E_\theta(X_t), Equation (2) gives Gθ(Xt,0)=U(t)+Dθ(Eθ(Xt))G_\theta(X_t,0)=U(t)+D_\theta(E_\theta(X_t)), which need not equal U(t)U(t). Equation (3) therefore does not enforce physical zero-time identity.

Likewise, latent composition does not establish alignment with an appropriately advanced history. The residual decoder remains anchored to the original observation, so latent composition is not equivalent to physical advancement followed by history re-encoding.

Increasing latent RMS cannot exclude sample-independent or other nonzero degenerate representations. The submitted diagnostics do not replace the decoded-consistency and alignment tests you request. The experiment concerns a particular latent penalty; it does not establish that physically consistent semigroup constraints are ineffective. Unregularized DT-FNO does not depend on a physical semigroup guarantee.

5. Reporting issues.

Table 4 is missing, and we apologize for the omission. Its caption cannot serve as quantitative evidence; we do not rely on those unreported physical metrics here.

We also acknowledge the inconsistent RealPDEBench summaries in Tables 2 and 9. We do not offer an unverified explanation or replacement value. These entries cannot support a precise estimate of the semigroup regularizer's effect without reconciliation.

The statement that DT-FNO outperforms R=16 at every tabulated horizon is incorrect. The Gray–Scott AR seed with one-step error 0.0136 outperforms DT's 0.0169. The reported advantage at h=32 and h=128 does not imply dominance at h=1, and the different DT training protocol remains a qualification.

Our defense rests on the explicitly identified empirical results and their limited scope, rather than missing metrics, unreconciled summaries, or a physical-consistency guarantee.

I thank the authors for the response and acknowledging the limitations. I do have several concers:

  • I thank the authors for their explanation and honest. However, the response does not address my concern; rather, it verifies it, as the authors explicitly acknowledge the limited methodological novelty.

  • As the contribution seems mainly controlled empirical comparisons. But from the manuscript, I can see DT introduces extra components and specific tuning for considered settings. And there is also no clear report of difference in training cost, potentially different hyperparameter choices, while differences in training costs and tuning budgets are not clearly reported. Given the raining-protocol difference, how can the authors attribute the improvements to the direct prediction instead of architectural or training differences?

  • I recommand that the authors provide the missing Table 4.

  1. Prior work and scope of the contribution.

Although we cited Poseidon, we did not adequately account for its time-conditioned FNO baseline (Appendix C.4) and its direct-versus-autoregressive comparison (Appendix D.6.2, Figure 47) when positioning our contribution. We apologize for this oversight. These experiments already include direct and autoregressive FNO inference and demonstrate task-dependent differences in their performance.

Accordingly, we do not claim priority for time-conditioned FNO or for the general idea of comparing direct and autoregressive forecasting. Our study extends this line of empirical investigation to history-conditioned forecasting with separately rollout-trained baselines, horizon-resolved evaluation, and backbone and climatology controls. The results characterize intermediate-horizon benefits and settings in which the advantage disappears or predictions approach climatology.

These differences specify the scope of our empirical contribution. They do not, by themselves, establish methodological novelty or resolve the attribution limitations discussed below. 2. Architectural controls, training differences, and attribution.

Several components mentioned in the review are shared by both models. Section 3.3 specifies the same history encoder, Fourier backbone, LayerScale residual blocks, and decoder. AR-FNO-R removes the time embedding and FiLM heads and is trained on its own multistep rollouts. The optional semigroup penalty is not used by the main DT-FNO model.

For Gray–Scott, the parameter counts are 8,414,786 for AR and 8,565,698 for DT; the difference is the 150,912 parameters in the time embedding and FiLM heads (approximately 1.8%). Appendix D.2 also reports common settings for width, depth, history length, optimizer, and learning-rate schedule, while explicitly listing dataset-dependent mode counts, padding, and horizon limits. These details specify the extent of matching; they do not establish identical architectures or equal tuning effort.

The width control in Appendix D.4 further shows that increasing the AR-FNO parameter count to approximately 15.9M does not reproduce the U-Net's accuracy and stability. This weakens a parameter-count-only explanation, but does not isolate the effects of time conditioning or temporal supervision.

Your attribution concern is therefore valid at the level of an isolated causal claim. Direct prediction changes both the conditioned model and its temporal training objective. Our experiments compare these forecasting configurations; they do not separate the accuracy contribution of avoiding recursive inference from that of horizon conditioning and multihorizon supervision. Accordingly, the supported claim is that the tested direct configuration achieves better intermediate-horizon field accuracy under the reported protocols. Attributing the entire improvement to direct inference alone would be stronger than our evidence permits.

Likewise, shared final hyperparameters do not establish equal hyperparameter-search budgets, and equal epoch counts do not establish equal training compute. Table 3 reuses the main-protocol DT result, so it cannot establish a fully compute-matched DT-versus-AR comparison or a precise end-to-end training-efficiency advantage. The existing controls narrow alternative explanations, but do not eliminate all architectural and training differences. We acknowledge this limitation explicitly.

One inference property can be attributed directly to the formulation: a single requested future state requires one DT evaluation instead of a sequence of AR steps. This evaluation-count advantage is separate from claims about accuracy, training cost, or tuning efficiency.

  1. Table 4.

Table 4 has now been included in the updated PDF. We apologize for the omission and thank you for identifying it.

Reviewer_KNFeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies the direct prediction of future physical states with a lead-time-conditioned FNO. Through experiments on five systems, it evaluates accuracy, inference cost, stability, and predictability limits in comparison with matched autoregressive models.

Critical strengths
  1. Direct temporal queries reduce inference cost and sidestep recursive error feedback.
  2. Matched baselines and targeted control experiments strengthen the comparison between direct and autoregressive prediction.
  3. Climatology comparisons help distinguish numerical stability from meaningful forecasting skill.
Critical weaknesses
  1. DTNO largely combines established lead-time conditioning and one-pass temporal querying with a standard FNO. Consequently, its contribution resides primarily in the controlled empirical comparison between direct and autoregressive prediction, rather than in a novel operator design.
  2. The accuracy advantage vanishes when U-Net is used on Gray–Scott, which restricts the generality of the main findings.
  3. Table 2 should clarify the meaning of entries marked "—", including the missing TpredT_{pred} values and horizon-specific errors, while Table 3 omits DT-FNO's training time and peak memory. Together, these omissions hinder a full assessment of predictive performance and training costs.
  4. While the semigroup regularizer improves the consistency of latent composition, it does not consistently enhance forecast accuracy; moreover, the paper does not examine whether direct and composed predictions remain consistent after decoding.
Questions for authors
  1. How exactly is TpredT_{\mathrm{pred}} defined and computed in Table 2, and what do the entries marked "—" denote? In particular, if the 0.3 threshold defined in Table 1 is likewise applied here, the reported errors for Lifted H₂ appear to have crossed this threshold. Could the authors therefore explain why TpredT_{\mathrm{pred}} remains unreported for this system?
  2. Table 3 caption indicates that the DT-FNO results were transferred from Table 2 rather than retrained within the rollout-depth sweep. Could the authors clarify which training settings differ between these two configurations, and, where feasible, report DT-FNO's training time and peak memory under a setting comparable to that of the other entries in Table 3?
2 comments

Thank you for the review and for recognizing the value of the matched comparisons and climatology controls. Our response clarifies what the existing results support and acknowledges the reporting gaps.

1. Empirical contribution and dependence on the backbone.

We agree that the contribution is primarily empirical. The architecture provides a lightweight implementation for studying direct and autoregressive forecasting under matched spatial backbones, rather than a new operator class.

The U-Net result is an important boundary of the finding. Appendix D.4 reports aggregate errors of 0.2952 ± 0.0191 for AR-U-Net and 0.2980 ± 0.0077 for the direct U-Net on Gray–Scott, so these experiments do not demonstrate a direct-model accuracy advantage. The claim is therefore not that direct forecasting universally dominates autoregression.

This result also makes the comparison informative: an approximately parameter-matched, wider FNO does not reproduce the U-Net's accuracy and stability, indicating that parameter count alone is insufficient to explain the difference. These controls support a backbone-dependent interpretation, without identifying a complete causal mechanism.

2. T_pred and the entries marked “—”.

The reporting is incomplete here. Table 1 explicitly defines a persistence-based threshold time using E_field=0.3. This dataset diagnostic must be distinguished from the model-specific threshold summaries in Table 2, whose computation is not fully documented.

Your observation about Lifted H₂ is correct: the reported errors bracket 0.3, so the missing values cannot be interpreted as showing that the threshold was never crossed. Likewise, the RealPDEBench DT-FNO error already exceeds 0.3 at the first reported horizon. An unexplained dash does not distinguish missing computation, an unevaluated horizon, or a threshold outside the evaluated range.

We acknowledge that this leaves your request for the exact computation and missing values unresolved in the current response. We do not introduce an unverified calculation convention or replacement value. The accuracy comparisons defended here use the explicitly reported horizon-specific field errors, not the missing threshold summaries. Such threshold times should also be understood as operational metrics, rather than intrinsic predictability limits.

3. Training settings and the missing DT-FNO costs.

The DT-FNO row in Table 3 was transferred from the main experiment, as stated in the caption. Appendix D.2 reports batch 64 per GPU on four GPUs for the main protocol, whereas Table 3 specifies batch 32 for the rollout-depth sweep. Equal epoch counts do not establish equal optimizer-update counts or equal compute budgets.

Accordingly, Table 3 is not a fully compute-matched comparison between DT-FNO and the rollout-depth arms. It provides measured training-time and memory scaling within the autoregressive sweep, together with a direct-model accuracy reference from the main protocol.

The separate fixed-batch microbenchmark in Section 5.3 supports the narrower observation that direct training has the same computational depth at h=1 and h=128. It cannot replace the missing end-to-end DT training-time and peak-memory measurements. We therefore do not infer a precise overall training-efficiency ratio from Table 3.

4. Interpretation of the semigroup regularizer.

We agree that improved latent composition consistency does not establish consistency after decoding or better forecast accuracy. The residual decoder also means that latent identity does not automatically enforce physical zero-time identity.

The reported evidence does not establish a consistent accuracy benefit from this particular latent penalty. This is not a general negative conclusion about physically meaningful semigroup constraints. The direct-versus-autoregressive comparison does not require the optional regularizer.

The empirical conclusion is conditional: the direct formulation offers a useful accuracy–cost alternative in the reported FNO settings, while the U-Net and climatology controls identify important limits to that advantage.

Thank you for the response and clarifications.

However, my concerns remain. The response confirms that the contribution is primarily empirical but does not resolve my concern regarding limited methodological novelty.

Moreover, the entries marked “—” in Tables 2 and 3 remain insufficiently explained. In particular, the response does not clarify why the TpredT_{\mathrm{pred}} values in Table 2 were not obtained or reported.

Since controlled empirical evaluation constitutes the main contribution, these unresolved reporting gaps affect my assessment of the completeness of the experimental evaluation.

Reviewer_VP2dOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies Direct-Time Neural Operators (DTNO), which predict a state at a requested future horizon in one evaluation rather than recursively applying a short-step model. Across five datasets, the matched autoregressive FNO is better at short horizons, while DTNO becomes better after approximately 8–20 steps and is substantially more stable across seeds.

Critical strengths
  1. Diverse evaluation includes reaction–diffusion, convection, wake flow, and reacting-flow systems.
  2. Reports complete error-versus-horizon curves, inference cost, divergence, seed variability, climatology baselines, and spectral diagnostics.
Critical weaknesses
  1. Novelty is limited: one-pass time-conditioned prediction already exists in Poseidon, TC-UNet, and DeepONets. The main contribution is therefore empirical, yet none of these direct-time methods is experimentally compared. Also, DTNO is an FNO with a Fourier time embedding, FiLM conditioning, and a residual decoder. These components and direct time-conditioned prediction already exist. The method does not introduce a new operator
  2. Under chaotic or partially predictable dynamics, minimizing squared field error produces a conditional-mean forecast. This can remain bounded by smoothing or averaging structures rather than predicting the correct realization. The climatology results suggest that this occurs on several datasets.
Questions for authors
  1. Does DTNO preserve trajectory-specific correlations, conservation laws, and temporal coherence beyond matching field error?
  2. What is the cost advantage when a dense trajectory, rather than one future state, is required?
2 comments

Thank you for the careful review. We would like to clarify the empirical contribution and distinguish the demonstrated forecasting benefits from properties that the current evidence does not establish.

1. Contribution and relationship to existing direct-time methods.

Our contribution does not depend on claiming that time conditioning, FiLM, FNO, or direct temporal querying is new. Sections 1–2 explicitly acknowledge the relevant prior work. We also recognize that Poseidon compares direct and autoregressive inference.

The distinction in our study is the comparison between a history-conditioned direct model and a separately rollout-trained autoregressive model, with matched spatial backbones, trajectory-level splits, and evaluation protocols. This differs from deploying the same trained time-conditioned model in two inference modes. The comparison concerns forecasting strategies, including their respective temporal training objectives, rather than inference alone.

The resulting evidence is more specific than a general claim that direct prediction is better: the tested FNO models exhibit an intermediate-horizon crossover; the advantage disappears with a U-Net on Gray–Scott; widening FNO does not reproduce the U-Net's stability; and time conditioning does not guarantee interpolation between sparsely supervised horizons. These findings help identify when the direct formulation is useful and when it is not. Comparisons with Poseidon, TC-UNet, and DeepONet would answer a broader benchmarking question that our current experiments do not settle. Our claims are therefore restricted to the comparisons actually reported.

2. Does the benefit reflect only convergence to a mean field?

We agree that bounded error alone is not forecasting skill. Section 5.5 explicitly makes this distinction, and Table 8 documents climatology-like spectral behavior on Lifted H₂.

However, the reported advantage is not limited to long horizons at which the autoregressive model diverges. At h=32, Table 2 reports DT/AR/trajectory-climatology field errors of 0.3683/0.5882/0.5639 on Gray–Scott and 0.2068/0.4523/0.6123 on Cylinder. These comparisons support useful intermediate-horizon field prediction in those settings. They do not establish accurate realization-specific prediction at arbitrarily long horizons.

The trajectory climatology is an oracle diagnostic using the test trajectory, not a deployable baseline or a rigorous lower bound. Conversely, approaching its error does not prove that the observed history contains no remaining predictive information.

3. Correlations, conservation, and temporal coherence.

The submitted field, spectral, and qualitative comparisons provide partial diagnostics, but they do not establish all the properties in your question. We do not claim exact conservation or guaranteed dynamical coherence across independently queried states. Section 6 explicitly identifies the latter limitation. The defensible accuracy claim concerns the reported field errors and horizons; it does not require interpreting numerical boundedness as physical fidelity.

4. Cost for dense trajectories.

For Q requested times with maximum horizon H, direct prediction requires Q state evaluations, while a one-step autoregressive model requires H evaluations and can retain all intermediate outputs. The evaluation-count advantage is therefore H/Q. If Q=H, both require O(H) evaluations.

Direct queries can be batched independently, but this does not establish a wall-clock advantage for dense trajectories without an appropriate benchmark. The reported constant-depth benefit applies to each requested future state. The practical case supported here is sparse querying at horizons where useful forecast accuracy is demonstrated.

I thank authors for the responses . However, after reading the author's response as well as other reviewers comment, my concern on the mythology novelty is still not addressed. Also, the author doesn't provide the results on more challenging case to demonstrate the behavior for long-term prediction and chaotic dataset where the traditional model might likely fail severely. Therefore, I'll maintain my recommendation on weak rejection.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.